Navigating the Legal Risks of Training AI Models on Copyrighted Books

Judge William Alsup's order imposing a $1.5 billion settlement on Anthropic penalized the company for sourcing books from illegal shadow libraries—not for the act of training itself. For teams building or fine-tuning language models, the precedent narrows the legal risk surface to data sourcing practices rather than the training pipeline.

Training ≠ Copying: The Alsup Framework

The ruling draws a functional analogy between LLM ingestion and human literary study. Alsup wrote that Anthropic's models trained on copyrighted works "not to race ahead and replicate or supplant them—but to turn a hard corner and create something different." This framing treats training as consumption rather than reproduction—a distinction copyright attorney Cathy Gellis considers broadly favorable for AI developers. Copyright law, she noted, hinges on copying, not on experiencing or reading a work.

The practical implication: the legal exposure vector shifts from "did you train on copyrighted data?" to "how did you obtain that data?" Licensed datasets, public-domain corpora, and properly acquired materials occupy a defensible position. Shadow libraries and pirated archives do not.

Fair Use Remains an Unresolved Variable

The 1976 Copyright Act provides no explicit guidance on machine learning. Courts evaluate fair use through four factors—purpose, nature, amount used, and market impact—but application to AI training remains inconsistent. Jason Henderson, founder of the IP & Media Practice at JWL International, described the judicial landscape as "all over the place" in its reasoning.

For ML teams, this means legal defensibility is fact-specific and jurisdiction-dependent. No blanket safe harbor exists. The $1.5 billion penalty against Anthropic—substantial in absolute terms but marginal relative to the company's projected $200 billion annual revenue by 2028—signals that courts may treat data sourcing violations as a cost-of-business calculation rather than an existential threat.

What Practitioners Should Track

Three variables warrant monitoring. First, pending litigation may produce conflicting rulings that further fragment the legal landscape. Second, legislative proposals to update copyright law for AI training could establish clearer rules—or impose new restrictions. Third, the distinction between licensed and unlicensed training data will likely become a standard due-diligence checkpoint, much like how developer tools define the infrastructure layer across adjacent technology sectors.

The core takeaway: the Alsup ruling does not greenlight indiscriminate data ingestion. It establishes that lawful acquisition is the operative constraint. Teams should audit their data pipelines accordingly.